Compound models and Pearson residuals for single-cell RNA-seq data without UMIs.

Area of Science:

• Computational Biology
• Genomics
• Statistical Modeling

Background:

• Existing methods for normalizing single-cell RNA sequencing (scRNA-seq) data often rely on unique molecular identifiers (UMI).
• Normalization of non-UMI data, such as Smart-seq2, presents challenges due to amplification biases and complex distributional patterns.

Purpose of the Study:

• To develop a novel statistical framework for normalizing non-UMI scRNA-seq data.
• To extend the utility of Pearson residuals for gene selection and dimensionality reduction to datasets lacking UMIs.
• To accurately model the technical noise inherent in non-UMI protocols.

Main Methods:

• Modeling sequenced RNA molecules using a negative binomial distribution.
• Incorporating an amplification distribution to account for technical biases in non-UMI data.
• Developing compound Pearson residuals based on the proposed compound distribution.
• Describing amplification distributions with a broken power law.

Main Results:

• The compound Pearson residual model effectively normalizes Smart-seq2 datasets.
• The model yields meaningful gene selection and informative embeddings for non-UMI data.
• A broken power law accurately describes amplification distributions across various sequencing protocols.
• The compound model successfully addresses overdispersion and zero-inflation patterns specific to non-UMI data.

Conclusions:

• The proposed compound distribution model offers a robust approach for normalizing non-UMI scRNA-seq data.
• This methodology enhances the interpretability and utility of gene expression data from protocols like Smart-seq2.
• The findings provide a more accurate statistical representation of RNA sequencing experiments without UMIs.